Multi-source sintering transfer learning in small dataset sintering prediction scenario

• DOI: 10.1007/s12289-021-01630-y
• Publication Date: 2021-06-16T08:03:02Z
• Authors: Wu Zhouzhi; Zhang Xiaomin; Zhao Zhipeng; Zhang Hengjia; Tang Hongwu
• Keywords: Benchmark (surveying); Transfer of learning

The purpose of this paper is to achieve fast densification prediction of target materials and particularly of difficult-to-sintering materials in small dataset sintering scenarios. A multi-source sintering transfer learning framework based on the domain adversarial network (DANN) was proposed to achieve multi-source sintering transfer learning. Further, a calibration method was presented to enhance the reliability of the integrated multi-source DANN (IMDANN), where DANN was applied as a test benchmark. The results indicate that IMDANN is significantly better than the test benchmark under all the test conditions. Error analysis illustrates that the root mean square error (RMSE) of IMDANN’s prediction converges to approximately 5% in the target domain, and the average prediction error is reduced by 56.5%. With an increasing number of source domains following the correlation criterion, the minimum number of source domains required for convergence is only 3–4. Compared to DANN, the calibrated IMDANN has high reliability and realises the cross-domain transfer of sintering knowledge with a small dataset.